DALL-E 2's Clever Trick: Predicting in CLIP Embedding Space, Not Pixels

thesephist · x · 2026-09-23

Linus Lee (thesephist) revisits a clever design in DALL-E 2: instead of predicting raw pixels or specialized latent patches, it generated in CLIP embedding space, enabling operations like interpolation in image space. He cites it as evidence that many unexplored generation paradigms remain.

Related event: Developer highlights DALL-E 2's clever CLIP embedding-space image generation(2 posts)→

Original post →

More from Multimodal

Multimodal channel →